maedhroz commented on code in PR #2267: URL: https://github.com/apache/cassandra/pull/2267#discussion_r1185330988
########## src/java/org/apache/cassandra/io/tries/IncrementalTrieWriterPageAware.java: ########## @@ -0,0 +1,464 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.cassandra.io.tries; + +import java.io.IOException; +import java.util.ArrayList; +import java.util.Comparator; +import java.util.List; +import java.util.NavigableSet; +import java.util.TreeSet; + +import javax.annotation.concurrent.NotThreadSafe; + +import org.apache.cassandra.io.util.DataOutputBuffer; +import org.apache.cassandra.io.util.DataOutputPlus; + +/** + * Incremental builders of on-disk tries which packs trie stages into disk cache pages. + * + * The incremental core is as in {@link IncrementalTrieWriterSimple}, which this augments by: + * <ul> + * <li> calculating branch sizes reflecting the amount of data that needs to be written to store the trie + * branch rooted at each node + * <li> delaying writing any part of a completed node until its branch size is above the page size + * <li> laying out (some of) its children branches (each smaller than a page) to be contained within a page + * <li> adjusting the branch size to reflect the fact that the children are now written (i.e. removing their size) + * </ul> + * <p> + * The process is bottom-up, i.e. pages are packed at the bottom and the root page is usually smaller. + * This may appear less efficient than a top-down process which puts more information in the top pages that + * tend to stay in cache, but in both cases performing a search will usually require an additional disk read + * for the leaf page. When we maximize the amount of relevant data that read brings by using the bottom-up + * process, we have practically the same efficiency with smaller intermediate page footprint, i.e. fewer data + * to keep in cache. + * <p> + * As an example, taking a sample page size fitting 4 nodes, a simple trie would be split like this: + * <pre> + * Node 0 | + * -a-> | Node 1 + * | -s-> Node 2 + * | -k-> Node 3 (payload 1) + * | -s-> Node 4 (payload 2) + * ----------------------------------- + * -b-> Node 5 | + * -a-> |Node 6 + * | -n-> Node 7 + * | -k-> Node 8 (payload 3) + * | -s-> Node 9 (payload 4) + * </pre> + * where lines denote page boundaries. + * <p> + * The process itself will start by adding "ask" which adds three nodes after the root to the stack. Adding "ass" + * completes Node 3, setting its branch a size of 1 and replaces it on the stack with Node 4. + * The step of adding "bank" starts by completing Node 4 (size 1), Node 2 (size 3), Node 1 (size 4), then adds 4 more + * nodes to the stack. Adding "banks" descends one more node. + * <p> + * The trie completion step completes nodes 9 (size 1), 8 (size 2), 7 (size 3), 6 (size 4), 5 (size 5). Since the size + * of node 5 is above the page size, the algorithm lays out its children. Nodes 6, 7, 8, 9 are written in order. The + * size of node 5 is now just the size of it individually, 1. The process continues with completing Node 0 (size 6). + * This is bigger than the page size, so some of its children need to be written. The algorithm takes the largest, + * Node 1, and lays it out with its children in the file. Node 0 now has an adjusted size of 2 which is below the + * page size, and we can continue the process. + * <p> + * Since this was the root of the trie, the current page is padded and the remaining nodes 0, 5 are written. + */ +@NotThreadSafe +public class IncrementalTrieWriterPageAware<VALUE> +extends IncrementalTrieWriterBase<VALUE, DataOutputPlus, IncrementalTrieWriterPageAware.Node<VALUE>> +implements IncrementalTrieWriter<VALUE> +{ + final int maxBytesPerPage; + + private final static Comparator<Node<?>> BRANCH_SIZE_COMPARATOR = (l, r) -> + { + // Smaller branches first. + int c = Integer.compare(l.branchSize + l.nodeSize, r.branchSize + r.nodeSize); + if (c != 0) + return c; + + // Then order by character, which serves several purposes: + // - enforces inequality to make sure equal sizes aren't treated as duplicates, + // - makes sure the item we use for comparison key comes greater than all equal-sized nodes, + // - orders equal sized items so that most recently processed (and potentially having closer children) comes + // last and is thus the first one picked for layout. + c = Integer.compare(l.transition, r.transition); + + assert c != 0 || l == r; + return c; + }; + + IncrementalTrieWriterPageAware(TrieSerializer<VALUE, ? super DataOutputPlus> trieSerializer, DataOutputPlus dest) + { + super(trieSerializer, dest, new Node<>((byte) 0)); + this.maxBytesPerPage = dest.maxBytesInPage(); + } + + @Override + public void reset() + { + reset(new Node<>((byte) 0)); + } + + @Override + Node<VALUE> performCompletion() throws IOException + { + Node<VALUE> root = super.performCompletion(); + + int actualSize = recalcTotalSize(root, dest.position()); + int bytesLeft = dest.bytesLeftInPage(); + if (actualSize > bytesLeft) + { + if (actualSize <= maxBytesPerPage) + { + dest.padToPageBoundary(); + bytesLeft = maxBytesPerPage; + // position changed, recalculate again + actualSize = recalcTotalSize(root, dest.position()); + } + + if (actualSize > bytesLeft) + { + // Still greater. Lay out children separately. + layoutChildren(root); + + // Pad if needed and place. + if (root.nodeSize > dest.bytesLeftInPage()) + { + dest.padToPageBoundary(); + // Recalculate again as pointer size may have changed, triggering assertion in writeRecursive. + recalcTotalSize(root, dest.position()); + } + } + } + + + root.finalizeWithPosition(write(root)); + return root; + } + + @Override + void complete(Node<VALUE> node) throws IOException + { + assert node.filePos == -1; + + int branchSize = 0; + for (Node<VALUE> child : node.children) + branchSize += child.branchSize + child.nodeSize; + + node.branchSize = branchSize; + + int nodeSize = serializer.sizeofNode(node, dest.position()); + if (nodeSize + branchSize < maxBytesPerPage) + { + // Good. This node and all children will (most probably) fit page. + node.nodeSize = nodeSize; + node.hasOutOfPageChildren = false; + node.hasOutOfPageInBranch = false; + + for (Node<VALUE> child : node.children) + if (child.filePos != -1) + node.hasOutOfPageChildren = true; + else if (child.hasOutOfPageChildren || child.hasOutOfPageInBranch) + node.hasOutOfPageInBranch = true; + + return; + } + + // Cannot fit. Lay out children; The current node will be marked as one with out-of-page children. + layoutChildren(node); + } + + private void layoutChildren(Node<VALUE> node) throws IOException + { + assert node.filePos == -1; + + NavigableSet<Node<VALUE>> children = node.getChildrenWithUnsetPosition(); + + int bytesLeft = dest.bytesLeftInPage(); + Node<VALUE> cmp = new Node<>(256); // goes after all equal-sized unplaced nodes (whose transition character is 0-255) + cmp.nodeSize = 0; + while (!children.isEmpty()) + { + cmp.branchSize = bytesLeft; + Node<VALUE> child = children.headSet(cmp, true).pollLast(); // grab biggest that could fit + if (child == null) + { + dest.padToPageBoundary(); + bytesLeft = maxBytesPerPage; + child = children.pollLast(); // just biggest + } + + assert child != null; + if (child.hasOutOfPageChildren || child.hasOutOfPageInBranch) + { + // We didn't know what size this branch will actually need to be, node's children may be far. + // We now know where we would place it, so let's reevaluate size. + int actualSize = recalcTotalSize(child, dest.position()); + if (actualSize > bytesLeft) + { + if (bytesLeft == maxBytesPerPage) + { + // Branch doesn't even fit in a page. + + // Note: In this situation we aren't actually making the best choice as the layout should have + // taken place at the child (which could have made the current parent small enough to fit). + // This is not trivial to fix but should be very rare. + + layoutChildren(child); + bytesLeft = dest.bytesLeftInPage(); + + assert (child.filePos == -1); + } + + // Doesn't fit, but that's probably because we don't have a full page. Put it back with the new + // size and retry when we do have enough space. + children.add(child); + continue; + } + } + + child.finalizeWithPosition(write(child)); + bytesLeft = dest.bytesLeftInPage(); + } + + // The sizing below will use the branch size, so make sure it's set. + node.branchSize = 0; + node.hasOutOfPageChildren = true; + node.hasOutOfPageInBranch = false; + node.nodeSize = serializer.sizeofNode(node, dest.position()); + } + + protected int recalcTotalSize(Node<VALUE> node, long nodePosition) throws IOException + { + if (node.hasOutOfPageInBranch) + { + int sz = 0; + for (Node<VALUE> child : node.children) + sz += recalcTotalSize(child, nodePosition + sz); + node.branchSize = sz; + } + + // The sizing below will use the branch size calculated above. Since that can change on out-of-page in branch, + // we need to recalculate the size if either flag is set. + if (node.hasOutOfPageChildren || node.hasOutOfPageInBranch) + node.nodeSize = serializer.sizeofNode(node, nodePosition + node.branchSize); + + return node.branchSize + node.nodeSize; + } + + protected long write(Node<VALUE> node) throws IOException + { + long nodePosition = dest.position(); + for (Node<VALUE> child : node.children) + if (child.filePos == -1) + child.filePos = write(child); + + nodePosition += node.branchSize; + assert dest.position() == nodePosition + : "Expected node position to be " + nodePosition + " but got " + dest.position() + " after writing children.\n" + dumpNode(node, dest.position()); + + serializer.write(dest, node, nodePosition); + + assert dest.position() == nodePosition + node.nodeSize + || dest.paddedPosition() == dest.position() // For PartitionIndexTest.testPointerGrowth where position may jump on page boundaries. + : "Expected node position to be " + (nodePosition + node.nodeSize) + " but got " + dest.position() + " after writing node, nodeSize " + node.nodeSize + ".\n" + dumpNode(node, nodePosition); + return nodePosition; + } + + protected String dumpNode(Node<VALUE> node, long nodePosition) + { + StringBuilder res = new StringBuilder(String.format("At %,d(%x) type %s child count %s nodeSize %,d branchSize %,d %s%s%n", + nodePosition, nodePosition, + TrieNode.typeFor(node, nodePosition), node.childCount(), node.nodeSize, node.branchSize, + node.hasOutOfPageChildren ? "C" : "", + node.hasOutOfPageInBranch ? "B" : "")); + for (Node<VALUE> child : node.children) + res.append(String.format("Child %2x at %,d(%x) type %s child count %s size %s nodeSize %,d branchSize %,d %s%s%n", + child.transition & 0xFF, + child.filePos, + child.filePos, + child.children != null ? TrieNode.typeFor(child, child.filePos) : "n/a", + child.children != null ? child.childCount() : "n/a", + child.children != null ? serializer.sizeofNode(child, child.filePos) : "n/a", + child.nodeSize, + child.branchSize, + child.hasOutOfPageChildren ? "C" : "", + child.hasOutOfPageInBranch ? "B" : "")); + + return res.toString(); + } + + @Override + public PartialTail makePartialRoot() throws IOException + { + // The expectation is that the partial tail will be in memory, so we don't bother with page-fitting. + // We could also send some completed children to disk, but that could make suboptimal layout choices, so we'd + // rather not. Just write anything not written yet to a buffer, from bottom to top, and we're done. + try (DataOutputBuffer buf = new DataOutputBuffer()) + { + PTail tail = new PTail(); + // Readers ask rebufferers for page-aligned positions, so make sure tail starts at one. + // "Padding" of the cutoff point may leave some unaddressable space in the constructed file view. + // Nothing will point to it, though, so that's fine. + tail.cutoff = dest.paddedPosition(); + tail.count = count; + tail.root = writePartial(stack.getFirst(), buf, tail.cutoff); + tail.tail = buf.asNewBuffer(); + return tail; + } + } + + protected long writePartial(Node<VALUE> node, DataOutputPlus dest, long baseOffset) throws IOException Review Comment: nit: @SuppressWarnings("DuplicatedCode") -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
